Robust Self-Augmented Open-Circuit Fault Diagnosis of Three-Level Inverters for EV Powertrains

Abstract

With the rapid advancement of data collection systems, data-driven fault diagnosis approaches have significantly improved the detection of open-circuit (OC) faults in voltage source inverters (VSIs). However, the diverse operating conditions and noisy environments in electric vehicles (EVs) cause feature variations, rendering challenges in maintaining the high accuracy and robustness of existing methods. This article introduces a robust self-augmented deep learning (DL) method for detecting OC faults in VSIs working at various operation conditions. The proposed method automatically extracts the most relevant statistical features from the measured signals using an automatic feature extraction module (AFEM) based on an improved particle swarm optimization (PSO) algorithm. Then, these optimized features are utilized through an attention-based multi-input convolution neural network (A-MICNN) embedded with a multihead self-attention module (SAM) to effectively capture the local and global characteristics of different OC faults. The A-MICNN architecture effectively processes these multi-inputs by employing shared weights and feature extraction layers to fuse information from statistical features and time-frequency image data. The proposed method is validated using an in-house three-level (3L) F-type inverter setup, achieving a remarkable diagnosis accuracy of 99.74% with high robustness across various speeds, load conditions, and noise levels, providing an efficient solution for EV powertrain fault diagnosis.